A dictionary-based method for crop yield prediction using satellite remote sensing data

HYDRA (HYbrid Dictionary–Rocket Architecture) has shown competitive performance across several benchmark datasets in prior Time-Series Classification (TSC) studies. It is a dictionary-based method that uses static convolutional kernels as dictionary patterns. HYDRA incorporates these non-learnable kernels in its method, which form the basis of ROCKET (RandOm Convolutional KErnel Transform) method designed for TSC tasks. This work presents a machine-learning-based framework built on HYDRA to perform regression on remotely sensed satellite data for crop yield prediction. It integrates HYDRA and crop phenological features through concatenation to generate a composite feature vector for XGBoost (eXtreme Gradient Boosting), enabling within-field corn yield prediction using time-series Sentinel-2 imagery. On the 2017 and 2018 datasets derived from a 22-ha corn field in Italy, our method achieved R2 values of 0.553 and 0.696, RMSE values of 1.101 t/ha and 1.013 t/ha, and MAE values of 0.796 t/ha and 0.708 t/ha, offering a modest improvement in yield prediction performance over frameworks based solely on HYDRA and phenological features. The application of HYDRA-based features in remote-sensing-based agricultural applications shows promising results for regression tasks.